> Source: https://mp.weixin.qq.com/s/EITf-SrP5o62Ljp7UGzPVw > Score: 81 · Category: ai-models · Published: 2026-07-21
At the 67th International Mathematical Olympiad (IMO 2026), held July 15–16 in Shanghai, 666 contestants from 117 countries competed, and only 7 human participants earned perfect-score gold medals. The dots team from Xiaohongshu (RED) was invited to participate: their internal model dots-note-3.0 scored a perfect 7/7 on all six problems, achieving 42/42.
Beyond the score itself, the team's technical write-up argues something bigger: recursive self-critique as a replacement for formal verification as a general engine for reasoning.
What It Did Right
First, full-process proofing without Lean.
Past "AI scores at IMO" results mostly took the formalized path: humans first translate problems into Lean or similar languages, then let models solve them. dots-note-3.0 instead received the organizers' original LaTeX problem statements and ran end-to-end in an agentic fashion — natural-language reasoning plus Python code. The formal path's limitation is that "many complex real-world problems cannot be fully formalized"; dots chose the harder but more general route.
Second, a Proof → Verify → Refine loop.
The model generates a candidate proof, checks it for correctness and completeness, revises the problematic parts, then submits. The Verify step is not mere right/wrong scoring — it identifies reasoning errors, logical jumps, and imprecisions, and feeds improvement suggestions into the next Refine round. This goes deeper than "voting + Best-of-N": it turns self-review into an optimization objective.
Third, parallel multi-round self-revision.
Rather than a single pass, the system runs "multi-round parallel reasoning, review, correction, and self-solution integration." The model runs several candidate proofs internally, uses Verify feedback to cross-correct them, and finally selects the strongest version.
Two Judge Comments Show Why This Time Is Different
Liu Hanzuo (two-time CMO gold medalist) wrote: "The model ultimately produced a correct and beautiful proof idea. The solution is compact in structure and natural in logic" — the emphasis is not just a correct answer but a *beautiful* solution. Wang Qiantong (CMO gold medalist), on problem 3: "The common approach among contestants is basically to convert this into graph connectivity, but the dots model used a very clever and elegant induction."
These comments indicate dots' solutions were not brute-force-searched correct answers, but proofs that competition-trained judges recognized as elegant, human-quality work. The model is approximating not just problem difficulty but the *taste* of human proof-writing.
Where This Road Leads
The dots team's write-up states: "We believe the premise of recursive self-improvement is recursive self-critique. Without actively identifying errors and deficiencies, real optimization is impossible; conversely, effective self-critique is what drives the model to approach correct answers."
Frankly, this is not a new idea. But dots embedded it into dots-note-3.0's training objective and produced a quantifiable outcome — a perfect IMO 2026 gold.
The deeper signal: most complex real-world tasks (scientific research, deep analysis, strategic planning, open-ended creation) cannot be fully formalized, and their results cannot be immediately verified by simple rules. In these fuzzy, non-automatically-scoreable domains, whether a model can continuously question and correct itself determines whether it moves from "guessing answers" to "doing research."
Details Worth Noting
- Model scale: note is the most lightweight model in the dots3 series; the full series includes jazz and aria (different sizes, response speeds, and compute costs). This means 42/42 was not "brute-forced with compute."
- Open-source commitment: dots expects to open-source note-3.0 "in the coming period." If delivered, this would be the first industrial-grade reference implementation in the "reasoning + self-critique" track among Chinese open-source LLMs in 2026 H2.
- No formalization dependency: Most industry reports of "AI perfect IMO scores" rely on formal tools like Lean. dots explicitly chose not to depend on human translation — an engineering-harder but academically more general path.
Caveats and What to Watch
Perfect IMO scores have appeared before (DeepMind, Google Research, OpenAI), so the raw score alone is not a scarce signal. But two things differentiate this run:
1. Not the formalization path — previous perfect scores relied on Lean translation; dots' version reads raw LaTeX directly. 2. Solutions judged "beautiful" by human evaluators — something a raw score cannot measure.
Open questions: Can third parties reproduce the same score once dots-note-3.0 is open-sourced? What is the specific training method for the Verify module? These will become clear after release.
Beyond the competition, the more important thing to track is whether the "self-critique-driven iteration" methodology transfers to non-mathematical tasks — research assistance, strategic games, complex code debugging. The dots team's stated position is clear: this will be a "core direction of continued exploration."
Source: Xiaohongshu Tech official WeChat account, 2026-07-21, "Xiaohongshu dots Model Achieves Perfect Gold at IMO 2026," by the dots studio team.